Few-shot learning with representative global prototype.

Area of Science:

• Machine Learning
• Computer Vision

Background:

• Few-shot learning models often struggle with generalization due to limited training data from base classes.
• Existing methods face challenges with data imbalance and prototype bias when incorporating novel classes.

Purpose of the Study:

• To propose a few-shot learning method that enhances generalization to novel classes.
• To address data imbalance and prototype bias in few-shot learning scenarios.

Main Methods:

• A strategy for jointly training base and novel classes to create representative global prototypes.
• A novel sample synthesis method to augment sparse novel class data.
• Selection of representative and uncertain samples to refine global prototypes.

Main Results:

• The proposed method significantly improves classification performance on few-shot learning tasks.
• Achieved state-of-the-art results on two popular benchmark datasets.

Conclusions:

• Joint training and sample synthesis effectively enhance few-shot learning generalization.
• The representative global prototype method offers a robust solution for few-shot classification challenges.